Master batch record vs batch record: where generative AI fits
GenAI can draft a master batch record or recipe change. FDA and draft Annex 22 say a qualified person must review and own it. What that review and the records look like.
Sources checked October 26, 2026. The US rules and the FDA letter are current. Annex 22 is a draft and the EMA workshop report is a meeting record, so neither is a requirement. Vendor figures are the vendor’s own claims.
Generative AI can draft a master batch record or a recipe change, but it cannot own one. FDA told one firm in an April 2026 warning letter that if AI helps create master production or control records, the firm must review them for accuracy and compliance. Draft EU Annex 22 says large language models (LLMs) should not be used in critical GMP applications. So the draft goes through the same change control, review and approval as any master record change.
Master batch record vs batch record, and can generative AI write one
A master batch record is the approved template for making a product. US rules call it the master production and control record. 21 CFR 211.186 says it must be “prepared, dated, and signed (full signature, handwritten) by one person and independently checked, dated, and signed by a second person”, and lists what it holds, from the formula to the in process controls. In a manufacturing execution system (MES), the master recipe plays the same role.
A batch record is the executed copy for one batch. 21 CFR 211.188 says it starts as “an accurate reproduction of the appropriate master production or control record” and records what actually happened. So an error in the master copies itself into every batch made from it.
That is the reason AI authoring of master records needs more care than AI review of executed ones. Both are being sold now. Körber announced on October 22, 2026 K.AI, built on Azure OpenAI and Azure AI Foundry, for digitizing, searching and updating recipes. The company says “Initial pilot projects indicate that recipes can be digitized around 30 percent faster and the proportion of manual steps reduced by up to 40 percent.” Those are pilot claims from the vendor. The Pharma AI Tracker manufacturing and quality area lists what drug companies themselves have disclosed in this space.
What FDA has already said
“If you use AI as an aid in document creation, you must review the AI generated documents to ensure they were accurate and actually compliant with CGMP. Your failure to do so is a violation of 21 CFR 211.22(c).”
FDA, warning letter to Purolea Cosmetics Lab, April 2, 2026
The same letter says the firm “used AI to create drug product specifications, procedures, and master production or control records”. I wrote about what the case does and does not show in One AI warning letter is not a trend, so I won’t retell it.
The two rules it leans on apply directly to master records. 211.22(c) gives the quality control unit “the responsibility for approving or rejecting all procedures or specifications impacting on the identity, strength, quality, and purity of the drug product.” 211.186 adds the two signature rule. An AI draft changes neither. It adds something new to check.
Master batch records under draft Annex 22: where LLMs stand
Draft Annex 22 section 1 says “the document does not apply to Generative AI and Large Language Models (LLM), and such models should not be used in critical GMP applications.” Critical means “direct impact on patient safety, product quality or data integrity”. In non critical uses, “personnel with adequate qualification and training should always be responsible for ensuring that the outputs from such models are suitable for the intended use”.
The draft gives no list of which uses count as critical. My view is that an LLM writing master record content with no human review would be critical by any reading, because the master sets the process for every batch. An LLM that drafts text which a qualified person then checks, edits and approves through change control is closer to the non critical case. Some sites will still treat it as critical, and that is a defensible call to write down.
The EMA workshop report (EMA/156789/2026, June 30, 2026) records industry arguing against a categorical ban, and lists among its examples “an LLM supporting GMP risk assessment for an electronic batch record system”. That is discussion, not a decision. How much testing each use needs follows its risk, which I cover in How much validation does your AI need.
What the AI may draft and what QA must own
Say a site lets an assistant draft master recipe changes from approved change requests. This is the split I would use.
| Task | AI may draft | QA must own |
|---|---|---|
| Digitizing a paper master | A first electronic version of the steps and fields | A line by line check against the approved paper master |
| Drafting a recipe change | Changed steps from an approved change request | Whether the change matches the request and the process |
| Parameters and limits | Nothing new. It may copy values from the approved source | Every value, checked against the registered process |
| Instruction wording | Clearer operator instructions | That the clearer wording means the same thing |
| Approval | A summary of what changed for the reviewers | The two signatures under 211.186 and release under change control |
I would not let an AI propose new process parameters at all. Limits come from development and registration data, not from text. For drafting help in GxP work more generally, see Can you use ChatGPT, Claude or Copilot for GxP work?.
The records that must exist
For an AI drafted master record change, I would expect five records. The change request that started it. The AI draft as produced, kept with the change. The human edits, as a compare against the draft. The preparer’s and checker’s signatures. The audit trail of the system that holds the master.
Current EU GMP Annex 11 already says changes to a computerized system “should only be made in a controlled manner in accordance with a defined procedure”, and expects audit trails for GMP relevant changes. The model call that produced the draft has its own log, which I described in an audit trail for LLM calls.
Is a hallucination check validation?
No. Körber says its PharmaGuardrails feature “checks every AI response using a specialized algorithm, helping to virtually eliminate AI hallucinations.” I have not seen data on that claim. Even if it works as described, it is a control inside the system. Validation is the site showing that the whole system, controls included, is fit for its intended use at that site, with that process and those records.
So I would test the check like any other control. Feed it drafts with known errors, such as a wrong limit, a missing step or a swapped unit, and record what it catches and what it misses. The misses tell you what the human review must look for.
If you are evaluating AI recipe authoring and want help with the review and records, get in touch.
Frequently asked questions
Can an LLM approve a recipe change?
No. 21 CFR 211.186 requires a master production and control record to be prepared and signed by one person and independently checked and signed by a second. 211.22(c) gives the quality control unit responsibility for approving procedures and specifications. FDA’s April 2026 warning letter puts the duty to review AI generated documents on the firm.
Is a master batch record a critical GMP application?
Draft Annex 22 does not list examples. It defines critical by direct impact on patient safety, product quality or data integrity. In my view, AI writing master record content without human review would be critical. AI drafting text that a qualified person checks and approves under change control is closer to non critical, but some sites treat it as critical, and that is a reasonable choice.
Does a vendor's hallucination check count as validation?
No. A hallucination check is a control in the system. Validation shows that the system as a whole, including that control, is fit for the site’s intended use. Test the check with drafts that contain known errors and record what it catches, then design the human review around what it misses.